I built a small dataset of subscription plan, team-seat and API token prices for ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google) and Grok (xAI), and put it on the Hub: geekindata/ai-plan-and-api-prices · Datasets at Hugging Face
How it works:
- Every row comes from the vendor’s own pricing or help page and carries the source URL, a status field and a verifiedOn date.
- A script re-reads those pages once a day. Numbers a vendor doesn’t publish are marked Not published, never stored as zero.
- API rates are per 1M tokens, with cached-input rates where the vendor lists them.
Files: plans (consumer plans with price, annual price and usage notes), team plans, and API prices, as CSV and JSON. The data is CC BY 4.0.
What it doesn’t do: it doesn’t rank model quality or speed, and it doesn’t cover Copilot, Perplexity or other apps.
The Hub snapshot is from 11 Nov 2026. The daily-updated files and a human-readable comparison are on the site I run, AI Price Data (Nov 2026): JSON & CSV, CC BY 4.0 , and the collection method is described on the site’s methodology page.
Feedback on the schema is welcome, especially if a field would make it easier to use in notebooks or eval cost estimates.
DHEIVER 2
@geekindata The distinction between full and cached input pricing is critical because it directly alters the marginal cost of RAG pipelines. When query redundancy is high, cached tokens drop the inference cost significantly compared to full generation. A dataset that separates these rates allows engineering teams to model total cost of ownership more accurately without assuming a uniform token price. Including a field for cache hit rates or a clear indicator of cached vs. non-cached rows would make cost estimations for on-prem deployments much more reliable.